Short answer
The cost of AI software development is build hours multiplied by the rate, plus ongoing model usage. At our €45–65 per hour, a proof of concept on your own data is about 2 engineers for 3 weeks: 240 hours, or €10,800–15,600. An AI feature inside an existing product is nearer €25,200–36,400, and an agent connected to your systems €59,400–85,800.
Two bills: building it and running it
The first bill is engineering time: preparing data, designing prompts and retrieval, connecting the model to your product, testing the answers and putting safeguards around them. It is priced like any software work, in hours at a rate.
The second bill arrives every month once people use it. Model providers charge per token, the units of text sent to and returned by the model, so cost follows volume and the length of each request. A design that sends less text per request can lower this bill for as long as the feature runs.
What drives the build hours
The model itself is often the smallest part of the effort. Hours usually go into:
- Data: finding, cleaning and structuring the documents or records the AI should use, and keeping them up to date.
- Retrieval: splitting content, creating embeddings and choosing a search approach so the model sees the right context.
- Evaluation: a set of real questions with expected answers, run on every change to measure quality instead of guessing.
- Guardrails: handling off-topic requests, sensitive data, wrong answers and when to pass the case to a person.
- Tool use: for agents, connecting safely to your CRM, ticketing, databases or internal APIs with the right permissions.
- Integration and interface: putting the feature into your product with sensible latency, streaming and error states.
- Privacy and compliance: GDPR, data residency and the provider's terms on how request data is stored and used.
Hosted model, fine-tuning or training your own
Most business AI products today call a hosted model through an API and add your data at request time through retrieval. This is the quickest and cheapest route to build, and you pay the provider for usage rather than for hardware.
Fine-tuning an existing model adds work to prepare training examples and evaluate the result, and makes sense when you need a consistent style or format that prompting cannot reach. Training a model from scratch, or hosting open models on your own GPUs, is a much larger commitment in data, infrastructure and specialist time, and is justified only by scale, strict data control or a capability no hosted model offers.
Model usage: the cost that scales with users
Estimate usage cost from three numbers: how many requests you expect per month, the average number of tokens in and out of each request, and your provider's current price per token. Long documents in the context, chat histories and agents that take several steps per task all multiply the token count.
Set budgets and alerts with the provider from day one, log token usage per feature, and review it alongside quality. Add vector database hosting, monitoring and the regular evaluation runs to the monthly figure.
Who builds it and at what rate
AI features need engineers who can ship production software as well as work with models, because most failures happen in data handling, integration and edge cases rather than in the model call. Our senior engineers work from Prishtina at €45–65 per engineer-hour under team extension.
For a complete AI product, a dedicated squad from €15k per month includes project management, QA and DevOps. A well-defined first version, such as an assistant over a fixed document set with agreed quality targets, can also be delivered at a fixed price, with MVPs starting from €15k.
Keeping AI costs under control
Both bills can be reduced without lowering the quality users see:
- Begin with a short proof of concept and an evaluation set before funding the full build.
- Use a hosted model and retrieval before considering fine-tuning or your own infrastructure.
- Route simple requests to smaller, cheaper models and reserve larger ones for hard cases.
- Send only the context a request needs, and cache answers to repeated questions.
- Keep a person in the loop for high-stakes outputs instead of engineering for perfection up front.
| Team and duration | Hours | Cost at €45–65/hour | |
|---|---|---|---|
| Proof of concept on your data, with an evaluation set | 2 engineers × 3 weeks | 240 h | €10,800–15,600 |
| AI feature in an existing product, such as search or summaries | 2 engineers × 7 weeks | 560 h | €25,200–36,400 |
| AI agent connected to your internal systems | 3 engineers × 11 weeks | 1,320 h | €59,400–85,800 |
| Fine-tuned model with a data pipeline and monitoring | 4 engineers × 18 weeks | 2,880 h | €129,600–187,200 |
